Skip to content

Dataset · lab260

Large Russian speech corpus for ASR with punctuation

lab260/balalaika-sft-rover7-punct

Nearly 12,000 hours of Russian speech with consensus transcriptions from seven ASR models and added punctuation.

What's inside

The dataset contains 4,775,837 Russian audio recordings with a total duration of 11,921.5 hours from 8 sources. For each audio, recognition variants from seven different ASR systems, the final consensus text with restored punctuation, the recording source, and the DistillMOS quality metric are saved. The exact file format is not specified in the card fragment.

What you can do with it

  • SFT and fine-tuning of Russian speech recognition (ASR) models with punctuation support
  • Training models for automatic punctuation and distilling ASR ensembles
Tasks
Automatic speech recognition
Languages
Russian
Size
1M–10M rows
Downloads this month
21.9K
Likes
3

How to load it

from datasets import load_dataset

ds = load_dataset("lab260/balalaika-sft-rover7-punct")